Parity doesn't, and was never intended to, decrease corruption.
All parity does is allow you to DETECT errors. If you detect errors, you still have to handle them.
Parity in memory is usually fatal - your code or data is corrupt, the solution is to halt the box, to stop this corruption (in memory) getting on to disk.
When used in communication (as it is here), parity errors usually mean you have to request a re-transmit (or, if acknowledgements are sent back, fail to send an acknowledgement and wait for an automatic re-transmit).
Also, be aware that the relevant error rate is per bit, not per packet. If a small packet (very small!) is 20 bytes, and there is corruption 0.5% of those packets, this indicates a 0.5% / 20 bytes per packet / 8 bits per byte (assuming 7 data bits, 1 parity bit, for example). This comes out to 0.003% errors.
Note that parity checks only work for 1-bit errors - in fact, a 2-bit error (in the same byte), the parity change will cancel out, and it will look ok.
An alternative to requesting re-transmits, is to use forward error correction - send redundant data (reducing your bandwidth). A very easy scheme is to send each packet twice (halving your bandwidth). You still need error checking (or some checksum) - if two packets are different, you still need to know which one was right. Checksums, on a more complex device, could let you try each combination, but for an Arduino, this would probably be too expensive (the code would take up sizable space, which an Arduino may, or may not, have spare).
All of this is usually written into higher level protocols (TCP, used for most internet file transmission, ZModem file transfer used for modems prior to the Internet). Other protocols, such as internet radio, will use either forward error correction, or will accept the occasional glitch.